Received 22 May 2003; received in revised form 18 November 2003; accepted 15 December 2003
Abstract
The performance of batch and semibatch reactors, under optimal operational conditions of amoxicillin (AN) enzymic synthesis at 25 ◦ C
and pH 6.5, were compared. Simulations for a batch reactor indicate that there are two sets of optimal operational conditions, if we consider
productivity and global yield (YAN,AB ) with respect to the ester substrate: (i) low initial concentration of ester and high initial concentration of
6-aminopenicillanic acid (6-APA); (ii) high initial concentration of ester and high initial concentration of 6-APA. Simulations for a semibatch
reactor, using the hybrid-neural network model, indicated a consumption of the formed antibiotic. Such pattern was not observed experimentally
and was caused by the lack of information during neural network training procedure. This problem was overcome by including new sets of
experimental data in the learning phase, performed in a batch reactor in the presence of different initial concentrations of substrates and
products. Results for the semibatch reactor indicated there are different sets of optimal operational conditions, if we consider local selectivity,
performance index (INH ), and global yield. Unfortunately, the point of highest local selectivity and performance index is not the one of highest
global yield. Therefore, the choice of initial operational conditions for the batch or semibatch production of amoxicillin is connected to an
economical evaluation. The semi-continuous operation of the integrated reactor (reaction + crystallization) provided better results than the
batch mode.
© 2004 Elsevier Ltd. All rights reserved.
Keywords: Pharmaceuticals; -Lactamic antibiotic; Penicillin G acylase; Simulation; Artificial neural network; Mathematical modeling; Immobilized enzyme
1. Introduction G acylase (PGA) could be of great interest due to its high se-
lectivity, specificity, and activity in mild reaction conditions
Amoxicillin (AN) is one of the major -lactam antibi- (aqueous media, neutral pH, and moderate temperatures).
otics, with sales of US$ 2200 million as a bulk formulated This work focuses on the production of amoxicillin, by
drug in 1994 [1], due to its high spectrum of activity, high the reaction of p-hydroxyphenylglycyne methyl ester (POH-
solubility, high rate of absorption, and its stability under acid PGME) and 6-APA catalyzed by penicillin G acylase immo-
conditions. Although industrial enzymatic synthesis is start- bilized on glyoxyl-agarose at 25 ◦ C and pH 6.5, see Fig. 1.
ing [2], today semisynthetic -lactam antibiotics are usually Previous works on the kinetics and mechanism of reaction
prepared in industry by chemical methods. These reactions [4–9] showed that the use of neural networks to represent
typically involve costly steps such as sub zero degree Celsius the kinetics seems to be an alternative to simulate experi-
conditions (−30 ◦ C) and toxic organic solvents like methy- mental data of amoxicillin synthesis, since the resulting set
lene chloride and silylation reagents [3]. Therefore, enzymic of mathematical equations of a deterministic model is very
synthesis is becoming increasingly interesting as an indus- complex. Hybrid models were used in this work to simu-
trial process because it reduces the number of reaction steps late the performance of batch and semibatch reactors for the
and decreases the amount and toxicity of waste products per production of amoxicillin.
kilogram of antibiotic. The use of enzymes such as penicillin Studies on the reactor optimization must be done, since
high product yield is required for the enzymatic synthesis to
∗ Corresponding author. Tel.: +55-85-2889601; fax: +55-85-2889601. become competitive with the chemical synthesis. Therefore,
E-mail address: lrg@ufc.br (L.R.B. Gonçalves). in this paper, the performance of a batch and a semibatch
0032-9592/$ – see front matter © 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.procbio.2003.12.010
248 L.R.B. Gonçalves et al. / Process Biochemistry 40 (2005) 247–256
O NH2 Synthesis O
CH3
S (vs) CH3
CH3 S
OH CH C OCH3 + OH CH C OCH NH CH3
N + CH3OH
COOH N
NH3 O NH3 COOH
O
p-hydroxyphenylglycine methyl ester 6-aminopenicilanic acid amoxicillin
Hidrolysis 1
+ H2O H2O +
(vh1) Hidrolysis 2
(vh2)
O O NH2 CH3
S
CH3
OH CH C OH + CH3OH OH CH C OH +
N
COOH
O
NH3 NH3
p-hydroxyphenylglycine p-hydroxyphenylglycine 6-aminopenicilanic acid
reactor were compared. For this purpose, techniques of pro- (bicarbonate buffer), in the presence of 100 mM of pheny-
cess modeling and simulation were used and the obtained lacetic acid (PAA), during 3 h, at 20 ◦ C. Final reduction of
results were compared to experimental data of synthesis, in the Schiff bases (reaction of the terminal and lysine amine
order to validate the mathematical model. groups of the enzyme with aldehyde groups in the support)
was obtained by adding sodium borohydride (1 mg/ml of
solution), during 30 min at room temperature. After each
2. Materials and methods step, the gel was filtered and washed with distilled water.
2.3. Mathematical models Two feedforward NNs were trained to represent the rates
of:
2.3.1. Semi-empirical kinetic model
The semi-empirical model that was used in this work is • amoxicillin net production (synthesis − hydrolysis) ⇒
described in detail elsewhere [4–6,12]. Shortly, its main νAN = νs − νh2 ;
hypotheses are: antibiotic synthesis only occurs when • p-hydroxyphenylglycine production (hydrolysis of the
6-aminopenicillanic acid (6-APA) binds to the enzyme be- ester + hydrolysis of amoxicillin) ⇒ νAOH = νh1 + νh2 .
fore the formation of the acyl-enzyme complex (EA) and
the rate of formation of the acyl-enzyme complex is not The experimental data used to train the neural network
influenced by the presence of 6-APA. Taking into account were obtained during amoxicillin synthesis at pH 6.5 and
those hypotheses, the rate equations become as follows: 25 ◦ C in a batch stirred tank reactor. During the learning
Rate of antibiotic hydrolysis (νh2 ): phase, the only output variable was the rate of reaction. The
kcat2 CE CAN
νh2 = (1)
Km2 (1 + (CAB /kAB ) + (CNH /kNH ) + (CAOH /kAOH )) + CAN
Rate of ester consumption (hydrolysis + synthesis, νAB ):
kcat1 CE CAB training procedure used here alternates the backpropaga-
νAB = tion algorithm [14] with a random-search method, aiming
Km1 (1 + (CAN /kAN ) + (CAOH /kAOH )) + CAB
at the minimization of the norm of the residues. For the
(2)
training/validation of the NNs we have used in-house soft-
Rate of amoxicillin synthesis (νS ): ware, from the Process Simulation Group of DEQ/UFSCar,
2.3.2. Hybrid-neural network model [5] 2.3.3. Mathematical model for the batch reactor
In this work, a feedforward neural network was used, Rates of reaction (semi-empirical or neural network ki-
trained by backpropagation, with information in one di- netic) were included in the batch reactor mass balances, as
rection only. The topology of the neural networks used in constitutive equations. Two mass balances unequivocally de-
this work is composed by four nodes in the input layer termine the system state (Eqs. (4) and (5)):
(substrates and products concentration) and one node in
dCAN
the output layer, related to the output variable (the kinetic = νAN (4)
rate), see Fig. 2. Only one hidden layer was used and the dt
number of nodes in this layer was determined by trial and dCAOH
error. = νAOH (5)
dt
250 L.R.B. Gonçalves et al. / Process Biochemistry 40 (2005) 247–256
W1
Amoxicillin
concentration
W2
p-hydroxypheyilglycine
concentration
6-APA
concentration rate of
reaction:
Ester AN or AOH
concentration
+1 +1
BIAS
Reaction stoichiometry provides the concentrations of the 2.3.5. Reactor performance indexes
remaining compounds. These ordinary differential equations Four parameters were defined to compare the reactors
could be easily solved by standard numerical methods, we performance:
have used Dassl subroutine [15].
• Selectivity (SAN ) in this work is the rate of desired
product (amoxicillin) per rate of undesired product
2.3.4. Mathematical model for the semibatch reactor
(p-hydroxiphenylglicine):
Rates of reaction (semi-empirical or neural network ki-
netic) were included in the semibatch reactor mass balances, νAN
SAN (%) = × 100.
as constitutive equations. Six mass balances determine the νAOH
system state (Eqs. (6)–(11)): • Performance index with respect to 6-APA (INH ) is the
dCAN,l amount of the desired product formed (amoxicillin) per
= νAN (6)
dt initial amount of 6-APA (substrate):
dCAOH,l CAN
= νAOH (7) INH (%) = × 100.
dt CNH0
dCAB • Global yield (YAN,AB ) with respect to the ester substrate
= −νAN − νAOH + FAB (8)
dt (AB), YAN,AB , is the amount of the desired product formed
dCNH (amoxicillin) per initial amount of POHPGME:
= −νAN + FNH (9)
dt CAN
YAN,AB (%) = × 100.
Substrates are added to the reactor in the form of powder, CAB0 − CAB
to avoid diluting the synthetic medium. Substrate were fed
• Amoxicillin productivity (PAN ) is the amount of the de-
separately into the reaction when its concentration reached
sired product formed (amoxicillin) per reaction time (tmax )
a minimum value, established at the beginning of the assay.
to achieve the highest amoxicillin concentration:
When antibiotic concentration reaches solubility
dCAN,l dCAN,s CAN
= 0 and = νAN (10) PAN =
dt dt tmax E0
When concentration of p-hydroxyphenylglycine reaches sol- where E0 is the reactor enzymatic load (IU/l).
ubility
dCAOH,l dCAOH,s
= 0 and = νAOH (11) 3. Results and discussion
dt dt
where l stands for liquid (soluble species) and s stands for 3.1. Batch reactor
solid (precipitated species). These ordinary differential equa-
tions were solved by standard numerical methods; we have Fig. 3 displays the obtained results of global yield with
also used Dassl subroutine [15]. respect to ester during amoxicillin synthesis at pH 6.5
L.R.B. Gonçalves et al. / Process Biochemistry 40 (2005) 247–256 251
80 CNH0 = 10 mM
)*100
80 70 CNH0 = 50 mM
CNH0 = 100 mM
YAN,AB (%) = CAN / (CAB0-CAB
60
60
50
40
40
30
20 20
100
80 10
0
30 60 30 40 50 60 70 80
40
)
CAB0 (mM)
M
50 40
(m
C 60
AB0 (mM 70 20
0
)
NH
80
respect to AB at pH 6.5 and 25 ◦ C. Different initial concentrations of
Fig. 3. Amoxicillin enzymatic synthesis, batch reactor: global yield with POHPGME and three different initial concentrations of 6-APA.
respect to AB (YAN,AB ) at pH 6.5 and 25 ◦ C.
40
30
INH (%)
20
10
0 100
30
80
40
60
50
C )
AB 60 40
mM
0 (m (
M) 70 0
20 C NH
80
Fig. 6. Amoxicillin enzymatic synthesis, batch reactor: performance index (INH ) at pH 6.5 and 25 ◦ C.
0.06 6
4
Concentration (g/l)
30
2
40
50 100 0
60 80
C
70
M)
40
Time (min)
M
(m
)
20
80 C NH0
Fig. 8. Validation test of the hybrid-neural network model for a semibatch
Fig. 7. Amoxicillin enzymatic synthesis, batch reactor: reactor productivity reactor: amoxicillin synthesis at pH 6.5 and 25 ◦ C with powder substrate
at pH 6.5 and 25 ◦ C. feed.
L.R.B. Gonçalves et al. / Process Biochemistry 40 (2005) 247–256 253
9
Soluble amoxicillin (Hybrid-neural network model)
8 Amoxicillin precipitated (Hybrid-neural network model)
Semi-empirical model
7
Concentration (g/l)
The hybrid-neural network model, however, failed in rep- thesis were performed, starting the reaction with different
resenting long reaction times. Fig. 9 shows the simulated re- concentrations of substrates and products. The neural net-
sults for amoxicillin synthesis at 25 ◦ C and pH 6.5, compar- work was trained again and new simulation results are pic-
ing the hybrid-neural network model and the semi-empirical tured in Fig. 10.
model for the semibatch reactor. It can be noticed that the simulated data no longer in-
It can be noticed that when reaction achieves 10 h, the hy- dicates the antibiotic consumption. For the same opera-
brid model indicated a consumption of the formed antibiotic, tional conditions, the semi-empirical semibatch model and
what was not expected (Fig. 9). This behavior is probable the hybrid-neural network semibatch model showed almost
caused by the lack of information during the learning phase the same behavior to amoxicillin synthesis at pH 6.5 and
of the neural network, used to represent the kinetics. The un- 25 ◦ C (Fig. 10).
expected pattern was not observed when the semi-empirical
kinetic model was used to simulate amoxicillin synthesis at 3.3. Semibatch reactor: simulation results
the same operational conditions, reinforcing the hypothesis
of problems during the learning phase of the neural network Based on the results obtained in this work and in previ-
kinetic model. New batch experiments of amoxicillin syn- ous works [5,6], the hybrid-neural network semibatch model
10
Time (min)
Fig. 10. Hybrid-neural network semibatch model validation test: amoxicillin synthesis at pH 6.5, 25 ◦ C.
254 L.R.B. Gonçalves et al. / Process Biochemistry 40 (2005) 247–256
80
YAN (%) = CAN /(CAB0-CAB )*100
60
40
20
80
70 0100
60 80
60
50
40
C AB
40
20 M)
(m
0
(m
30 0
M
C NH0
)
80 40 synthesis
100 30 C AB0
0
(m
M
Fig. 12. Amoxicillin enzymatic synthesis, fed-batch reactor: performance formance index obtained to both reactors studied in this
index (6-APA) at pH 6.5 and 25 ◦ C. work for the production of amoxicillin. Fig. 15 shows the
L.R.B. Gonçalves et al. / Process Biochemistry 40 (2005) 247–256 255
Table 2
Global yield YAN,AB and performance index INH for batch or semibatch 10 Batch Reactor
production of amoxicillin at 25 ◦ C and pH 6.5 Fedbatch Reactor
Parameter Batch reactor Semibatch reactor 8 Mathematical model
Global yield, YAN,AB (%) 82.0 82.0
Performance index, INH (%) 30.0 38.0 6
CAN (g/l)
0
1,5
SAN = vAN / vAOH
Time (min)
1,0 Fig. 16. Amount of amoxicillin produced per volume of a batch and a
semibatch reactor at pH 6.5 and 25 ◦ C.
4. Conclusion
0,5
A hybrid-NN model was able to capture the complex ki-
0 200 400
netics of the enzymatic synthesis of amoxicillin. Simulated
Time (min)
productivity and yield surfaces indicate that cost/profit ob-
jective functions are not unimodal. Optimum reactor opera-
Fig. 14. Amoxicillin enzymatic synthesis, fed-batch reactor at 25 ◦ C and tional strategies should be put forth with care. Simulations
pH 6.5: local selectivity time course. for a batch reactor indicated that there are two sets of opti-
mal operational conditions, if we consider global yield and
amount of amoxicillin produced per volume of a batch and productivity. The point of highest YAN,AB is not the one
a semibatch reactor. of highest productivity. The point of highest YAN,AB is in-
It can be observed that, although, selectivity and yields deed the point of minimum productivity. The hybrid-neural
of amoxicillin are quite similar to both reactors, the amount network semibatch model is able to represent experimental
of amoxicillin produced per volume of reactor is higher to data, when the set of experimental data, used in the learn-
the semibatch reactor. The semibatch reactor presented also ing phase, are obtained from batch assays in the presence
another advantage: while the ester substrate was fed to the of different initial concentration of substrates and products.
reactor, there was no antibiotic hydrolysis (Fig. 16 ). Simulations for the semibatch reactor indicated there are dif-
ferent sets of optimal operational conditions, if we consider
local selectivity a performance index with respect to 6-APA
CAB0 = 50 mM and CNH0 = 80 mM consumption and yield. Notice that the point of highest lo-
0,06
cal selectivity (SAN ) was not the one of highest global yield.
Therefore, the choice of initial operational conditions for the
0,05 batch or semibatch production of amoxicillin is connected
to an economical evaluation. Last but not least, the perfor-
mance of batch and semibatch reactors were compared on
PAN (g/IU min)
0,04
the point of optimal global yield and it was observed that the
0,03
semibatch reactor had a better performance than the batch
reactor, when considering amount of amoxicillin produced
per volume of reactor. As a role, the semi-continuous op-
0,02
eration of the integrated reactor (reaction + crystallization)
provided better results than the batch mode.
0,01
0 200 400
Acknowledgements
Time (min)
Fig. 15. Amoxicillin enzymatic synthesis, fed-batch reactor at 25 ◦ C and The authors would like to thank Hispanagar S.A. and
pH 6.5: productivity time course. Antibióticos S.A. for the agarose gel and the enzyme; and
256 L.R.B. Gonçalves et al. / Process Biochemistry 40 (2005) 247–256
the Brazilian research-funding agencies FAPESP (State of [2] Moody H, Hogenboom A, Lange B, Heemskerk D, Dooren TV,
Sao Paulo), CNPq and CAPES (Federal). Boesten W, et al. Enzymatic production of cephadroxyl. In: Pro-
ceedings of the 10th European Congress on Biotechnology, CAT 12,
2001.
Appendix A. Nomenclature [3] Ospina S, Barzana E, Ramı́rez OT, López-Munguı́a A. Effect of pH
in the synthesis of ampicillin by penicillin acylase. Enzyme Microb
Technol 1996;19:462–9.
AB p-hydroxyphenylglycine methyl ester [4] Gonçalves LRB, Fernandez-Lafuente R, Guisán JM, Giordano RLC.
AN amoxicillin A kinetic study of the synthesis of amoxicillin using penicillin G
AOH p-hydroxyphenylglycine acylase immobilized on agarose. Appl Biochem Biotechnol 2000;84–
6-APA 6-aminopenicillanic acid 86:931–45.
Ci concentration of i [5] Gonçalves LRB, Fernandez-Lafuente R, Guisán JM, Giordano RC,
Giordano RLC. Modeling the synthesis of amoxicillin: getting around
E enzyme limitations of the mechanistic approach for enzymatic kinetics.
FAB feeding of solid POHPGME per reactor Biotechnol Bioeng 2002a;80:622–31.
volume (mM/min) [6] Gonçalves LRB, Fernandez-Lafuente R, Guisán JM, Giordano RLC.
FNH feeding of solid 6-APA per reactor The role of 6-aminopenicillanic acid on the kinetics of the enzy-
volume (mM/min) matic synthesis of amoxicillin catalyzed by penicillin g acylase im-
mobilized on glyoxyl-agarose. Enzyme Microb Tech 2002;31:464–
IU international unit 71.
kcat kinetic rate constant, mechanistic model [7] Kasche V. Mechanism and yields in enzyme catalyzed equilibrium
ki inhibition constant and kinetically controlled synthesis of -lactam antibiotics. Enzyme
KEN 6-APA adsorption constant Microb Technol 1986;8:4–16.
NN neural network [8] Kasche V, Haufler U, Zöllner R. Kinetic studies on the mech-
anism of the penicillin amidase-catalyzed synthesis of ampicillin
PAA phenylacetic acid and benzylpenicillin. Hoppe-Seyler’s Physiol Chem 1984;365:1435–
PDAB p-dimetilaminobenzaldehide 43.
PGA penicillin G acylase [9] Kato K. Kinetics of acyl transfer by ␣-amino acid ester hydrolase
POHPGME p-hydroxyphenylglycine methyl ester from Xanthomonas citri. Agric Biol Chem 1980;44:1083–8.
Tmax maximum conversion ratio of the complex [10] Guisan JM. Aldehyde–agarose gels as activated supports for
immobilization-stabilization of enzymes. Enzyme Microb Technol
acyl-enzyme-nucleus into product 1988;10:375–82.
νAB rate of POHPGME consumption (mM/min) [11] Fernandez-Lafuente R, Rossel CM, Guisan JM. The use of stabi-
νAN rate of amoxicillin production (mM/min) lized penicillin acylase derivatives improves the design of kinetically
νAOH rate of POHPG production (mM/min) controlled synthesis. J Mol Catal A 1995;101:91–7.
νh1 rate of POHPGME hydrolysis (mM/min) [12] Gonçalves LRB, Fernandez-Lafuente R, Guisán JM, Giordano RC,
Giordano RLC. Inhibitory effects in the side reactions occurring dur-
νh2 rate of amoxicillin hydrolysis (mM/min) ing the enzymatic synthesis of amoxicillin: p-hydroxyphenylglycine
νS rate of amoxicillin synthesis (mM/min) methyl ester and amoxicillin hydrolysis. Biotechnol Appl Biochem
W vector of weights of the NN 2003;38:77–85.
X fraction of enzyme saturated with 6-APA [13] Marquardt DW. J Soc Ind Appl Math 1963;11:431–41.
[14] Rumelhart DE, Hinton GE, Willians RJ. Learning representations by
back-propagating errors. Nature 1986;323:533–6.
References [15] Petzold LR. Dassl D Code, Version 1989. Computing and Mathe-
matics Research Division, Lawrence Livermore National Laboratory,
[1] Diender MB, Straathof AJJ, Van der Wielen LAMS, Ras C, Heijnen 1989.
JJ. Feasibility of the thermodynamic controlled synthesis of amoxi- [16] Fogler, HS. Elements of Chemical Reaction Engineering. 3rd ed.
cillin. J Mol Catal A 1998;5:249–53. New York: Prentice Hall; 1999.